How Mnemor works

Last updated: 30 July 2026

Most apps that touch your mail say "we use AI" and leave you to guess what that means. This page is the opposite: every workflow in Mnemor, step by step. Any step that can send text off your phone is marked as one — the rest, silently, cannot.

The short version. Deciding what needs you is ordinary code running on your iPhone — no model, no network. AI is only used for things you explicitly ask for, like a drafted reply or a summary. Most of the steps below never leave the device.

Reading the diagrams

Each workflow is a list of steps. The dot beside a step tells you the only thing that really matters about it:

Stays on your phone
On-device where the hardware allows
Sends one excerpt out
Waits for your tap

Anything that can leave the phone is also labelled in words, so you never have to rely on the colour. An unlabelled step did not send anything.

What actually runs a step

Four kinds of machinery, plus you. Only one of the four can reach the internet.

Layer	What it is	Leaves the phone?
Rules	Deterministic Swift running on the device — scoring, matching, parsing, thresholds. No model is involved at any point.	No.
On-device	Apple's frameworks: on-device text recognition (Vision), on-device speech recognition, and Apple's on-device language model on supported hardware.	No.
Cache	A previous result, stored encrypted on the device and reused so the same question is not asked twice.	No.
Cloud AI	One text excerpt sent to Anthropic — either with your own API key or the built-in one. Logged in the app every time.	Yes — the excerpt only, and only in this layer.
You	A human tap. Nothing is sent, approved, replied to or merged without one.	Whatever you chose to send.

The one path to a model

There is exactly one function in the app that can talk to a language model. Every feature goes through it, so the controls below cannot be bypassed by any individual feature — including one we add later.

Mnemor's AI settings, with an option to use your own Anthropic key and toggles for each agent. — Where you choose your key, your model, and which agents may run at all.

Seven checks, in this order. Three of them can end the request before anything leaves your phone.
Subscription check

If you are in read-only mode, the call stops here and never happens.
Prompt-injection hardening

Instructions are added to every call so that text arriving from an email, a web page or a pull request cannot redirect the assistant.
Credential scrubbing

API keys, bearer tokens, JWTs, private-key blocks, explicit password: lines and Luhn-valid card numbers are stripped out before the text reaches any prompt — including an on-device one. This is hygiene, not anonymisation; it does not remove names or companies.
The hard stop

If On-device AI only is switched on, the request may use Apple's on-device model and nothing else. If that model is unavailable, the call returns empty. It never falls through to the cloud, for any feature, ever.
On-device first

Short, frequent work — suggestions, thread summaries, drafts, rewrites — is answered by Apple's on-device model where the hardware supports it, and stops here.
Your key, or ours Leaves the phone

Only now does anything go out. If you supplied an Anthropic key it is used, so the traffic runs under your own account and agreement; otherwise the built-in key is. Either way the request goes from your phone straight to Anthropic — never through a Mnemor server.
Written to your log

The time, the destination and the approximate size. The content never is.

There is no image path. That function accepts text and nothing else — no image, audio or file parameter exists anywhere in the app. Photos, documents and meeting audio therefore cannot reach a model, whatever the settings say. It is structural, not a preference.

Workflows

1. Your morning list

The core loop — the thing you actually open the app for.

Mnemor's ranked list of items that need a reply. — The result: one short list, ranked, with no model involved.

Every step stays on your phone. There is no AI anywhere in this workflow.
Fetch

Your phone calls Gmail, Outlook, Slack, GitHub, Calendar and Reminders directly, over TLS, using the tokens you authorised. Nothing is proxied through us.
Work out what is an obligation

Unanswered messages, reviews blocking someone, invitations awaiting a reply, things due today — identified by rules, not by a model.
Score and rank

Who has waited longest, what is due soonest, what blocks other people, what you have historically acted on. Money and security flags add weight. All of it is arithmetic on your device.
Protect the list

Advisory items are capped so they can never crowd out a real obligation. A newsletter cannot outrank a person.
Five things, each with a reason

Every item can tell you why it ranked where it did — because the ranking is explainable code, not a model's opinion.

2. Answering something

The first workflow where AI can appear — and only after you open the item.

A Mnemor card opened, showing the message, what Mnemor read, and a Send reply button labelled 'This sends for real'. — The card you tapped. The button says it plainly: this sends for real.

Nothing is sent until you open the item. Both AI steps prefer Apple's on-device model.
You open a card

Nothing was sent anywhere before this moment.
Already summarised?

If this thread was summarised before, the stored result is reused and no call is made.
Summarise the thread Cloud only as fallback

The recent messages of that one thread — not your mailbox — are condensed. Apple's on-device model does this where available; otherwise this is a cloud call, and it is logged.
Draft a reply Cloud only as fallback

Suggested actions and a reply, from the summary above. They are offered only when there is real context to act on — never guessed from a bare subject line.
You edit and send

Mnemor never sends a message on its own. Every reply, approval, RSVP and merge needs your tap — including from the lock screen.

3. Drafted before you even look

The background agents. Opt-in, throttled, and they never send.

Mnemor's follow-up settings, with an interval picker and a control for how many items to show. — Both agents are opt-in, and you set how patient they are.

Opt-in and throttled. It drafts, and then it stops — it cannot send.
Pick a few worth preparing

At most five items, no more than once every few hours, and only if you turned the agent on. Selection is rule-based.
Draft it Cloud only as fallback

On-device model first, cloud as fallback. For a follow-up, all that is sent is the subject and how long you have been waiting.
Sanity-check the output

A deterministic check discards refusals and malformed or over-long output rather than storing it.
It waits for you

The draft sits on the card until you open it. Nothing has been sent, and nothing will be.

4. Recording a meeting

Audio never leaves the phone. Not in any mode, not with any setting.

Mnemor's This week screen, with meeting recording, travel wallet and an end-of-day wrap up. — Recording starts here. The transcript is produced on the phone.

The audio never leaves your phone. One step sends the text transcript; switch it off and you keep the raw transcript.
Transcribe as you go

Apple's on-device speech recognition. The audio is never uploaded and is not retained after the transcript is produced.
You stop the recording

The microphone is released immediately, including if you just swipe the screen away.
Turn the transcript into notes Leaves the phone

The text transcript is summarised into decisions and action items. This is the one cloud step, and On-device only mode suppresses it — you keep the raw transcript.
Your action items join the list

What you committed to in the room becomes tracked items, ranked with everything else.

5. Photographing a receipt, badge or boarding pass

The whole workflow is on-device. No model is involved at any step.

Mnemor's scan screen for reading a conference badge or business card with the camera. — Camera in, text out — the image itself has nowhere to go.

Every step stays on your phone. No model is involved — this workflow works with the cloud switched off.
Read the image

Apple's Vision text recognition extracts the text on the device. The image itself is never transmitted — there is nowhere in the app for it to go.
Pull out the fields

Merchant, total and date from a receipt; name, company and title from a badge; flight, gate, seat and date from a boarding pass. Pattern matching, not AI.
Stored, sealed

Written encrypted on the device. Travel documents are additionally kept out of device backups and locked behind Face ID.

6. Nothing dropped

Catching your own promises. Entirely rule-based.

Mnemor's notification settings, showing daily nudges, routines and wind-down. — The morning heads-up and the end-of-day sweep are scheduled here.

Every step stays on your phone. Pattern matching, not AI.
Hear the promise

Phrasing like "I'll send it Friday" in something you wrote is detected by pattern matching on the device.
Track both directions

What you owe other people, and what you are still waiting on from them — with the wait time ageing in the ranking.
The end-of-day sweep

Three short lists before you log off: owed, awaiting, promised. Swipe to clear.

7. Asking a question

Search, by text or voice, across everything you have connected.

Mnemor's search screen, letting you ask a question across every connected source. — Ask in words or by voice, across everything you connected.

Your corpus is never uploaded. One step sends your question plus the items that matched it.
Hear the question

Voice input is transcribed on-device.
Find candidates locally

Your items, notes and meetings are matched on the device. The corpus is never uploaded.
Compose the answer Leaves the phone

Your question plus the titles and snippets that matched are sent — the matches only, not everything you have.
An answer that cites its items

You can open every item it drew on.

Every AI call the app can make

The complete list. Nothing else in the app calls a model. Fast uses Apple's on-device model where the hardware supports it, and only reaches the cloud otherwise; Smart is a cloud call, and stands down entirely in On-device only mode.

What you tapped	Tier	What is sent
Suggested actions on a card	Fast	That item plus its thread summary
Thread summary	Fast	Recent messages of that one thread
The brief's opening line	Fast	Titles of the current five (has a non-AI fallback)
A pre-drafted reply	Fast	Thread summary and what you decided before
A follow-up nudge draft	Fast	Subject and how long you have waited
Draft intro to a new contact	Fast	Name, company, role
Polish / shorter / warmer / firmer	Fast	Only the text you typed
Explain this	Smart	That item's title and context
Quick review of a pull request	Smart	The diff
Search or voice ask	Smart	Your question and the items that matched
Catch me up	Smart	Titles of what changed while you were away
Weekly review	Smart	Your on-device decision log for the week
Brief me on a person or project	Smart	Items and decisions mentioning them
AI spend narrative	Smart	Spend figures and money-flagged titles
Your world digest	Smart	Public headlines filtered to your industries
Meeting notes	Smart	The on-device transcript text
Prep me for this meeting	Smart	Attendees and what is open with them
Evaluate a capture	Smart	Text read from it, plus your question
Summarise a shared document	Smart	Text extracted from it
Handle this	Smart	Thread summary and your free slots
Research a contact you scanned	Smart	The only step that searches the web. Name and company only — and it is disabled unless both are known, so a bare name is never guessed at.

Why the same question is rarely asked twice

Results are cached on the device, encrypted, and thrown away when you change model. Reopening a card does not repeat a call.

Result	Reused for
Thread summaries, suggestions, explanations	Until evicted
Pre-drafted replies	24 hours
Person or project brief	1 day
Spend narrative	3 days
Weekly review, world digest	6 days
The brief's opening line	Until the five actually change

Checking any of this yourself

Open Settings → Privacy & security → Activity log. Every call on this page appears there with its time, its destination — on-device or a named provider — and roughly how much text was sent. Content is never recorded. Review the last 3, 7 or 30 days.

To end all of it in one switch: Settings → Privacy & security → On-device AI only. The workflows above marked Rules and On-device carry on exactly as they are; the cloud steps stand down rather than degrade quietly. The full security model · what each connector reads.